Constitutional AI turns safety rules into a structured generate, critique, revise loop, so principle compliance becomes an inspectable step instead of a hope.
Imagine writing an essay. One approach is to read a list of rules first and then write, hoping you followed them. Another approach is to write a draft, then re-read the rules and mark every place your draft broke one, then rewrite. The second approach catches problems the first never sees, because checking is a separate step you actually do. Constitutional AI is the second approach for an LLM. Instead of just listing rules in the system prompt, you ask the model to draft an answer, critique it against named principles, and revise. That extra step turns a vague instruction into an inspectable workflow.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Constitutional AI lives in two worlds at once. It is the name of a 2022 Anthropic training method. It is also the name of a prompting pattern that ships in production guardrail stacks today. The two share the same core idea, but they live at very different layers of the stack.
The prompting pattern is what most LLM engineers deploy.
The training story vs the prompting story
The 2022 paper introduced a training pipeline with two phases. In the supervised phase, the model critiques and revises its own responses against a written set of principles. Then the original model is fine-tuned on the revised responses. In the RL phase, a model judges which of two samples is better, a preference model is trained on those AI preferences, and RL runs against it. That second phase is called RLAIF, reinforcement learning from AI feedback. The critique-and-revise loop is a training-data step: the trained model does not run it at inference.
The prompting story is different. It uses the same loop structure (generate, critique, revise) but runs it at inference time, on any model, with no weight changes. Inference-time self-critique of this kind was studied separately, for example as Self-Refine (Madaan et al. 2023). The loop is a prompt scaffold plus an orchestration layer that calls the model two or three times per turn.
The common confusion comes from conflating these. The claim that Constitutional AI requires fine-tuning is wrong about the prompting pattern. The claim that the paper's models run the loop at inference is wrong about the training story.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Self-Refine (Madaan et al. 2023) runs generate, feedback, and refine with a single model at inference time, with no extra training, and reports about 20 percent average absolute gains across seven tasks.
- NVIDIA NeMo Guardrails ships self-check rails, where a model checks the user input or the bot output against a policy prompt and blocks it if it fails.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you write the principles list for a customer-support assistant?
Frame as 3 to 7 named, testable principles tied to support policy; each one phrased as something the critique step can answer yes or no on.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Confusing the Constitutional AI training method (from the 2022 Anthropic paper) with the prompting pattern. The inference-time pattern is just generate, critique, revise against named principles, and any model can run it.
60 second bullets to scan on the way to the call.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.